Inquiry Regarding MixMHC2pred Output and PWM Calculation

[winnieWei123456]

Dear developers of MixMHC2pred,

I hope this message finds you well. I am writing to seek clarification on the output of the first block of MixMHC2pred. In your article "Machine learning predictions of MHC-II specificities reveal alternative binding mode of class II epitopes," you mentioned that the output of the first block is the PPM. However, when reviewing the results, I noticed that the PPM contains values greater than 1. Shouldn't the sum of the frequencies of the 20 amino acids at each position be equal to 1?

Furthermore, I also took a look at your PWMdef file and noted the presence of 'PWM_norm'. What does 'norm' mean?Considering your statement in the F.A.Q section on the Motif Atlas website that "A correct description of the MHC binding specificity needs to account for this bias by renormalizing the amino acid frequencies computed in the raw ligands," I am curious to know if this PWM matrix has undergone frequency normalization. Could you please explain how this frequency normalization process is conducted?

I am interested in obtaining the original PWM. Could you provide guidance on how to calculate this?

I appreciate your time and assistance in addressing these queries. Thank you for your attention to this matter.

Warm regards,
Winnie

[jracle85]

Hello Winnie,

The "PWM" files that we use in our predictor correspond to

$$PWM^{a,s}_{l,i} = \frac{PPM^{a,s}_{l,i}}{f_i}$$

(as in the Equation 2 from our 2023 Immunity paper); where the PPMs are the position probability matrices and $f_i$ is the frequency of amino acid $i$ in the human proteome, used to normalize these PPMs. The PPMs will sum to 1 and not the PWMs. We used the following human proteome frequencies:
A=0.0693, C=0.0218, D=0.0483, E=0.0721, F=0.0355, G=0.0651, H=0.0258, I=0.0434, K=0.0577, L=0.0983, M=0.0219, N=0.036, P=0.0632, Q=0.048, R=0.0565, S=0.084, T=0.0545, V=0.0602, W=0.0122, Y=0.0261.

Best regards,

Julien